Who measured the AI cyber agents’ results?
Of the results I have on what AI cyber agents can do, only DARPA's contest had an outside referee; the rest are vendors reporting on themselves. The review of 14 October keeps the two kinds apart, because who did the measuring changes what a number can support.
One of the two things that started this research was reading about AI agents behaving outside their intended boundaries. The July Hugging Face incident was meant to be covered by the 00:40 review on 14 October, which was never saved in full, and the 01:41 review does not reconstruct it. So this research does not yet hold a checked account of that incident, and this piece does not give one.
The contest found 54 of 63 planted flaws and patched 43
DARPA's AI Cyber Challenge was externally administered. In the controlled setting, systems discovered 54 of 63 synthetic vulnerabilities (flaws made for the contest) and patched 43. On real software they found 18 vulnerabilities and submitted 11 patches.
DARPA corrected an earlier denominator of 70 to 63, and the review treats 63 as the authoritative figure. It adds that competition success does not mean every patch is correct under deployment conditions.
The vendor figures are reports on themselves
Anthropic's 6 October disclosure reports at least 129,000 partner-verified findings between April and July 2026, and another 5,500 from its own open-source scanning between April and October. The review lists what limits the reading: the partner data is partial (33 partners), triage criteria differ between them, patch counts are incomplete, and the review did no independent verification. The figures are reported observations; they do not show that 134,500 unique exploitable vulnerabilities were reproduced or fixed.
In the same disclosure a tiered trial shows the access-control trade-off: in the Defense tier 46 of 50 attempts were blocked at some point, and in the Red Team tier 34 of 50 were completed.
OpenAI's 1 September assessment of Astra reports exploit chains found and stronger monitoring and containment. It is an internal evaluation, and the cited test configuration is not the default user deployment.
Still open: evidence that a patch is correct
The review asks whether an agent's output can come with machine-checkable evidence of patch correctness, authorisation scope and noninterference (that it changed nothing outside what it was allowed to touch). A benchmark score is not a formal guarantee of safe behaviour.
It proposes an experiment. On a permissioned corpus of synthetic vulnerabilities, require proof obligations, regression tests, fuzzing (feeding a program random inputs to make it fail) and reproducible builds before a generated patch is accepted. Then compare that verified-fix rate with the superficial benchmark pass rate, including false positives and negatives.
Nothing has been run, so it remains a hypothesis.
Sources and notes (5)
- DARPA: AI Cyber Challenge results — 54 of 63 found, 43 patched
- Anthropic: Expanding the Cyber Verification Program — a vendor's report on itself
- OpenAI: Path to Astra — a vendor's internal evaluation
- Industry review, 14 Oct 01:41 — the notes this was written from
- Run log — the notes this was written from
Cryptography researchAll 10 parts →
